Neural Network Modeling for Beef Sensory Evaluation

Abstract

Feedforward backpropagation neural network models were developed for predicting and classifying beef sensory attributes using ultrasonic spectral features as input data. For the neural network prediction models, the standard errors of prediction were 0.126 for juiciness, 0.111 for muscle fiber tenderness, 0.102 for connective tissue amount, 0.113 for overall tenderness, and 0.135 for flavor intensity. These results were better than those of the statistical regression models. In the case of the neural models for beef sensory prediction, the relationships between physical (ultrasonic) features and sensory for beef were mostly linear. The neural network models for classification of the sensory attributes into two classes (threshold = 5.0 sensory score) performed with 83.3% accuracy for juiciness, 80.5% for connective tissue amount, 75.0% for muscle fiber tenderness and overall tenderness, and 75.0% for flavor intensity.